A bioengineered immunocompetent human leukemia chip for preclinical screening of CAR T cell immunotherapy

Chimeric antigen receptor (CAR) T cell immunotherapy is promising for treatment of blood cancers; however, clinical benefits remain unpredictable, necessitating development of optimal CAR T cell products. Unfortunately, current preclinical evaluation platforms are inadequate due to their limited physiological relevance to humans. We herein engineered an organotypic immunocompetent chip that recapitulates microarchitectural and pathophysiological characteristics of human leukemia bone marrow stromal and immune niches for CAR T cell therapy modeling. This leukemia chip empowered real-time spatiotemporal monitoring of CAR T cell functionality, including T cell extravasation, recognition of leukemia, immune activation, cytotoxicity, and killing. We next on-chip modelled and mapped different responses post CAR T cell therapy, i.e., remission, resistance, and relapse as observed clinically and identify factors that potentially drive therapeutic failure. Finally, we developed a matrix-based analytical and integrative index to demarcate functional performance of CAR T cells with different CAR designs and generations produced from healthy donors and patients. Together, our chip introduces an enabling ‘(pre-)clinical-trial-on-chip’ tool for CAR T cell development, which may translate to personalized therapies and improved clinical decision-making.

physiological relevance to humans. We herein engineered an organotypic immunocompetent chip that 23 recapitulates microarchitectural and pathophysiological characteristics of human leukemia bone marrow 24 stromal and immune niches for CAR T cell therapy modeling. This leukemia chip empowered real-time 25 spatiotemporal monitoring of CAR T cell functionality, including T cell extravasation, recognition of 26 leukemia, immune activation, cytotoxicity, and killing. We next on-chip modelled and mapped different 27 responses post CAR T cell therapy, i.e., remission, resistance, and relapse as observed clinically and 28 identify factors that potentially drive therapeutic failure. Finally, we developed a matrix-based analytical 29 and integrative index to demarcate functional performance of CAR T cells with different CAR designs 30 and generations produced from healthy donors and patients. Together, our chip introduces an enabling 31 '(pre-)clinical-trial-on-chip' tool for CAR T cell development, which may translate to personalized 32 therapies and improved clinical decision-making.

36
Adoptive chimeric antigen receptor (CAR) T cell transfer has emerged as an encouraging immunotherapy 37 for relapsed and refractory hematological malignancies, such as B cell acute lymphoblastic leukemia (B-38 ALL) [1][2][3], Diffuse large B cell lymphoma (DLBCL) [4,5], and multiple myeloma (MM) [6][7][8]. However, 39 therapeutic outcomes vary across clinical trials, for example nearly half of patients yield to disease relapse 40 with years of follow-up besides the side effects of cytokine release syndrome (CRS) and immune effector 41 cell-associated neurotoxicity syndrome (ICANS) [9][10][11]. Several mechanisms including T cell dysfunction 42 [12][13][14], surface antigen loss, down-regulation, and mutation [15][16][17], lineage swift [18,19] Current strategies for preclinical assessment of CAR T cell function rely on in vitro and in vivo models 49 [21][22][23]. Conventional in vitro assays such as two-dimensional (2D) cell co-cultures and 3D tumor 50 spheroids/organoids demonstrate limited predictive value due to their lack of tumor associated stroma 51 and/or immune components, thus are just useful for CAR T cell development at early-stage. In vivo animal 52 models have been widely established for preclinical cancer research, however, most of them are either 53 immunodeficient or differing from host immunity, dictating the pursuit of humanized models [21][22][23][24]. 54 Critically, current humanized animal models are painstaking due to extensive preparation and subsequent 55 month-long experiments and thwart real-time and in situ monitoring of CAR T cell response [9]. It thus 56 demands a reliable precision immuno-oncology tool that enables rapid, in-depth evaluation of CAR T cell 57 therapy within a human pathophysiologically relevant context [25][26][27]. In this study, we developed an ex vivo organotypic and immunocompetent human leukemia 60 microphysiological system exemplifying a bond fide leukemia bone marrow niche integrative of both 61 stroma and immune cells. Through on-chip real-time live cell imaging, comprehensive proteomic and 62 secretomic profiling, and high-throughput single cell mRNA sequencing (scRNA-seq), we precisely 63 captured systematical and spatiotemporal dynamics of CAR T cell treatment, ranging from T cell 64 extravasation, immune activation, and T cell cytotoxicity to T cell killing. Also, our bioengineered chip 65 accurately validated functional performance of CAR T cell products with different designs and generations 66 of CAR and those produced from cancer patients. Lastly, we developed a matrix-based analytical and 67 functional index to comprehensively delineate optimal CAR T cell. This unique preclinical platform 68 enables a precise, reliable, and systematic evaluation of CAR T cell therapy, which can be readily extended 69 to evaluate many other immunotherapies for different blood cancers as well as solid tumors and beyond.

72
Engineering an ex vivo immunocompetent human leukemia chip 73 To fill the biological and technical gaps in delivering an in-depth evaluation of CAR T cell therapy 74 dynamics within a human relevant context, we established an ex vivo vascularized and immunocompetent 75 human leukemia bone marrow niche on-chip. Anatomically, leukemia bone marrow is a multi-76 compartmental milieu divided into regions of central sinus, medullary cavity, and endosteum, where 77 dysfunctionally orchestrated interactions between leukemia blasts and hematopoietic and non-78 hematopoietic niche cells in 3D extracellular matrices (ECMs) maintain disease progression and promote 79 therapy resistance (Fig. 1a, left). To replicate the in vivo tissue architecture in vitro, our microfluidic chip 80 is designed with three interconnected structural regions, including a central sinus linked to concentric 81 medullary cavity encircled by outer endosteal region (Fig. 1a, middle and Extended Data Fig. 1a,b) and 82 fabricated using replica molding [28][29][30]. The chip was populated with human hematopoietic cells (bone 83 5 marrow mononuclear cells), leukemia blasts, stromal cells (vascular cells, mesenchymal stem cells, 84 osteoblasts, and fibroblasts) within compartmentalized fibrin hydrogels inside which the seeded bone 85 marrow stromal cells rebuilt and supported the three concentric tissue structures (Methods). We next 86 exploited the de novo vascularization process of vascular cells under guidance from intrinsic and extrinsic 87 cues such as vascular growth factors and ECMs to vascularize the tissue structure [31][32][33]. In brief, the 88 cell-laden chip was cultured with a myriad of cytokines (maintaining hematopoietic cells and promoting 89 vasculogenesis) over 7 days during which vascular cells self-assembled into perfusable vascular network 90 aligned by stromal cells throughout the 3D matrix (Fig. 1a, right and Extended Data Fig. 1c-h). 91 Hematopoietic cells and leukemia blasts sparsely distributed within the peri-/vascular and endosteal 92 niches, which well matched the in vivo cellular localization and composition (Extended Data Fig. 1d-h). 93 Notably, bone marrow stromal cells (mostly vascular cells) progressively deposited supplementary ECMs 94 such as laminin, fibronectin, and collagen IV around the vasculature and throughout the niche (Extended 95 Data Fig. 2), highlighting cell-autonomous remodeling of engineered tissue into a more biomimetic one 96 [34][35][36]. To validate the biomimicry of bioengineered bone marrow niche to its in vivo counterpart, we 97 next utilized scRNA-seq technique to comparatively map cellularity in the sample of cells harvested from 98 our chip and matched sample of unloaded fresh bone marrow mononuclear cells (Fig. 1b, Fig. 3, and Methods). The scRNA-seq results validated that our bone marrow chip, even after over 100 7-day culture, maintained an enriched cellularity with hematopoietic and stromal cells, comparable to that 101 of the in vivo bone marrow (Fig. 1d), keeping with recent scRNA-seq studies [28,[37][38][39][40]. We further 102 corroborated this observation with immunostaining and found most types of lymphoid (e.g., CD8 + and 103 CD4 + T cells) and myeloid (e.g., CD14 + monocytes and CD68 + macrophages) cells were well maintained 104 on-chip along with the presence of vascular network ( Fig. 1e-g and Extended Data Fig. 1). Together, our 105 tissue-engineered leukemia chip recreates in vitro microarchitectural organization and pathophysiological 106 signature of both human leukemia bone marrow stroma and immune niches in vivo.
Modeling CAR T cell therapy on-chip 108 We then modeled anti-CD19 CAR T cell therapy on-chip within a human pathophysiologically relevant 109 context and monitored CAR T cell dynamics in a real-time manner (Fig. 2, Extended Data Fig. 4, 110 Extended Data Fig. 5, and Methods). We first infused 10,000 (effector:tumor = 1:1) of second generation 111 (2nd-gen) anti-CD19 4-1BBζ-CAR T cells (hereafter CAR T cell, unless stated otherwise) per chip into 112 the perfusable vessels from central sinus and quantified the count of pre-seeded leukemia blasts on-chip 113 (GFP-expressing Reh, unless stated otherwise) daily for over 7 days. The results showed CAR T cell killed 114 ~70% of leukemia blasts on-chip after 3 days and achieved complete eradication (>99%) of leukemia 115 blasts around 7 days, reminiscent of clinical remission, whereas leukemia chips either treated with non-116 transduced T cell (referred as to Mock T cell) or left untreated (referred as to None) showed no control 117 over leukemia progression (Fig. 2a). Then, we longitudinally charted the migratory behaviors of CAR T 118 cell (labeled with DiD dye) in the medullary cavity region (vessel formed by RFP-expressing HUVECs) 119 for 12 hours with fluorescent live imaging after infusion for 2-4 days ( Fig. 2b and Extended Data Fig.   120 4a-c). We found that CAR T cells actively patrolled the leukemia bone marrow niche such as extravasated  (Fig. 2c). This reduced velocity of CAR T cell may be due largely to its recognition of CD19 + leukemia 129 blasts and contact with the latter to form an intercellular synapse (Extended Data Fig. 4d), where cytolytic 130 granules for example are released. Examining CAR T cell functionality on-chip 132 We next characterized CAR T cell functionality, such as T cell activation, cytotoxicity, and proliferation, 133 after its interaction with leukemia blasts in niche on-chip for 2 days. Compared to Mock T cell, CAR T 134 cell significantly enhanced surface expression of T cell activation markers, CD25 and CD69 (Fig. 2d,e 135 and Extended Data Fig. 5a-f), secretion of cytotoxicity-related cytokines, interferon-γ (IFN-γ, Fig. 2f we observed an excessive on-chip production of cytokines, such as immune stimulatory cytokines (e.g., 147 GM-CSF), chemokines (e.g., RANTES), inflammatory cytokines (e.g., MCP-2), and regulatory cytokines 148 (e.g., Interleukin(IL)-10), among many others (Extended Data Fig. 6c). This augmentation of cytokine 149 secretion presented only in chips where CAR T cell interacted with CD19 + leukemia blasts but not in those 150 where Mock T cell interacted with CD19 + leukemia blasts or CAR T cell with CD19 knockout (CD19 -) 151 ones (Extended Data Fig. 6c), confirming the specificity and efficacy of anti-CD19 CAR T cell therapy. Following this, we applied scRNA-seq technique to comparatively dissect the leukemia bone marrow 154 niches treated with CAR T cell or Mock T cell on-chip for 2 days (Fig. 2h, Extended Data Fig. 3, and 155 8 Extended Data Fig. 7). First, scRNA-seq results reaffirmed reduced population of CD19 expressing 156 clusters in niche upon treatment with CAR T cell but not Mock T cell (Fig. 2i). We found that CAR T 157 cell, compared to Mock T cell, enhanced mRNA expression of T cell activation and proliferation related 158 genes such as IL2RA, CD69, and MKI67 ( Fig. 2j and Extended Data Fig. 7a-c), validating 159 aforementioned observations with on-chip immunostaining. Moreover, activated CAR T cell increased 160 mRNA expression of cytotoxicity related genes such as GZMB and GNLY (Fig. 2k) and immune response 161 signaling pathways such as T cell receptor signaling pathway, Th1 and Th2 cell differentiation, Natural 162 Killer cell mediated cytotoxicity, and leukocyte transendothelial migration (Fig. 2l) Mapping heterogeneous CAR T cell response scenarios 173 Our understanding of the dynamic changes in the leukemia bone marrow microenvironment during CAR 174 T cell therapy responses such as resistance or effective killing are unclear, which may prevent unleashing 175 its full therapeutic potential [42,43]. We herein modelled the processes of leukemia resistance and relapse 176 after CAR T cell therapy using our bioengineered leukemia chip and mapped the underlying molecular 177 and cellular changes (Fig. 3). Clinical diagnosis by cytomorphology sets leukemia burden in bone marrow 178 below 5% as remission while above 25% as relapse or resistance after treatment, though other assays such 179 9 as PCR and flow cytometry take a lower percentage (0.01%-0.1%) as remission [44][45][46][47]. We here followed 180 the standard set by cytomorphology as it could allow us to quantify the number of leukemia blasts in a 181 single chip continuously during CAR T cell therapy with time-lapse imaging. To on-chip replicate relapse 182 scenario from pre-existing minor CD19clones, we spiked CD19leukemia blasts (5% or 1%) into the 183 total leukemia population when preparing leukemia chips, while to mimic resistance (i.e., refractory) 184 scenario, we infused 2,500 CAR T cells (effector:tumor = 1:4) per chip compared to 10,000 in the 185 remission and relapse scenarios (Methods). Then, we monitored leukemia burden chronologically by 186 fluorescence imaging of on-chip CD19 + leukemia blasts with GFP signal and CD19ones with mCherry 187 signal every day for over 14 days. The on-chip response curves showed that 2,500 CAR T cells mostly  (Fig. 3e). To identify molecular changes across different response scenarios, we further monitored 205 dynamics of cytokine secretion (i.e., IFN-γ) and found that IFN-γ was highest in remission scenarios 206 compared to that in resistance and relapse ones (Fig. 3f). These results together reproduced clinical 207 resistance and relapse scenarios, offering us an experimental model to real-time dissect CAR T cell therapy 208 within the whole disease spectrum.

210
Validating functional performance of CAR T cell products 211 We next leveraged our leukemia chip to preclinically assess the therapeutic potency of CAR T cell prod-212 ucts generated with different designs/generations and from both healthy donors and cancer patients. We 213 first validated our recently developed 2nd-gen anti-CD19 CAR T cells with different designs of CAR, i.e., 214 CD28ζ-CAR, ICOSζ-CAR, and 4-1BBζ-CAR, and found that these CAR T cells all achieved remission 215 on-chip at a dose of 10,000 CAR T cells (Extended Data Fig. 10a). Also, these CAR T cells enhanced 216 surface expression of CD25 and CD69, though at different levels, upon activation in niche on-chip, com-217 pared to that of mock T cell (Extended Data Fig. 10b). Intriguingly, cytokine profiling from on-chip tests 218 demonstrated that 4-1BBζ-CAR secreted more IL-13 (a Th2 cytokine), and CD28ζ-CAR and ICOSζ-CAR 219 secreted more IL-10, whereas this is not the case for 2D tests with respective co-cultures of CAR T cell We then assessed the functionality of 3rd-gen CAR T cell product where along with co-stimulatory signal 226 4-1BB, another functional domains such as IL-18 is integrated either to potentiate T cell activation or 227 11 mobilize the bone marrow immunity [52]. On-chip analyses confirmed that 3rd-gen 4-1BBζ-CAR-IL18 228 achieved a rapid remission response at high dose of 10,000 CAR T cells (Fig. 4a). To reveal functional 229 difference across different CAR designs, we run a 'CAR stress test' with low ratios of effector to tumor 230 (2,500 or 1,250 CAR T cells per chip), adapting our recent protocol [53]. The results showed that 2nd-gen 231 4-1BBζ-CAR took a longer time to eradicate leukemia blasts than did 3rd-gen 4-1BBζ-CAR-IL18 and 232 this difference was further expanded at a lower dose ( Fig. 4a and Extended Data Fig. 12). Next, we 233 profiled cytokine secretion from the 3rd-gen CAR T cell and normalized it to that of 2nd-gen CAR T cell. 234 We found that 4-1BBζ-CAR-IL18 demonstrated a robust cytokine secretion upon activation in niche on-235 chip ( Fig. 4b), confirming the role of integration of IL18 to strengthen CAR T cell potency. To better 236 outline the cytokine secretion of CAR T cell, we classified these cytokines into five categories, i.e., effec-237 tor, stimulatory, chemoattractive, inflammatory, and regulatory, according to its function in immune re-238 sponse processes and quantified the overall performance using weighted average. Again, 4-1BBζ-CAR-239 IL18 boosted cytokine secretion in all the categories, including those Th2 cytokines ( Fig. 4c and Extended 240 Data Fig. 12). This may partially explain why 2nd-gen 4-1BBζ-CAR underperformed 3rd-gen 4-1BBζ-241 CAR-IL18 during CAR stress test. 242 243 Lastly, we evaluated 2nd-gen anti-CD19 CAR T cell products generated from four patients either with B-244 ALL leukemia (PD323, PD356, and PD674) or non-B-ALL cancer (PD145) using our bioengineered chip 245 (Extended Data Fig. 13a,b). Of these four patients, on-chip remission was only achieved with treatment 246 of PD145 CAR T cells that enhanced surface expression of CD25 at day 2 ( Fig. 4d and Extended Data 247 Fig. 13c,d). Also, PD323, PD356 and PD674 CAR T cells demonstrated a significantly weakened cyto-248 kine secretion index compared to that of PD145 CAR T cell, benchmarked by a HD CAR T cell (Fig. 4e,   249 Extended Data Fig. 6c, and Extended Data Fig. 13e). The varied outcomes can be partially explained 250 by the limited expansion of T cells from PD323, PD356, and PD674 during CAR T cell manufacturing as 251 12 on-chip treatments were all given at a dose of 10, 000 CAR T cells (Extended Data Fig. 13b). To better 252 correlate on-chip performance with future clinical outcome, we developed a matrix-based analytical and 253 integrative index to systematically delineate the therapeutic potency of patient CAR T cells where PD145 254 CAR T cell outperformed other PD CAR T cells (Fig. 4f). Together, our bioengineered leukemia chip 255 demonstrated a useful and powerful platform for CAR T cell development.

371
To prepare the leukemia chip model, human leukemia blasts, bone marrow stromal cells, and bone marrow 372 immune cells were loaded into the microfluidic chips with physiologically relevant seeding density for 373 each cell type (i.e., HUVECs at 1×10 7 cells/mL, hMSCs at 1×10 5 cells/mL, NHLF at 2×10 6 cells/mL, 374 hMSC-derived osteoblasts at 2×10 6 cells/mL, Reh B-ALL at 1×10 6 cells/mL, and bone marrow 375 mononuclear cells at 5×10 6 cells/mL) in a fibrin hydrogel. In general, a multiple-step loading protocol 376 were followed to compartmentalize HUVECs in the center of the chip, HUVECs, hMSCs, Reh B-ALL, 377 and bone marrow mononuclear cells in the perivascular area, and NHLF and hMSC-derived osteoblasts 378 in the endosteal region. Unless stated otherwise, CD19 expressing K562 leukemia cell lines were used 379 18 with a seeding density at 5×10 5 cells/mL for 3rd-gen CAR T cell testing experiments. First, a sacrificial 380 gelatin (Cat#G6144-100G, Sigma) hydrogel solution of 12 mg/mL in phosphate buffered saline (PBS,381 Cat#10010049, Thermo Fisher Scientific) was injected into the central area and solidified at -20°C for 15 382 minutes. This step aims to minimize the generation of bubbles between different regions during the 383 following cell loading process. Then, a mixture of HUVECs, hMSCs, Reh B-ALL cells, and bone marrow 384 mononuclear cells in fibrin solution (3 mg/mL in PBS) containing 2 U/mL thrombin (Cat#604980-100U, 385 Sigma) was infused into the inner ring area and gelled at room temperature for 10 minutes. To recreate 386 the endosteal niche, a mixture of NHLF and hMSC-derived osteoblast in fibrin solution (3 mg/mL in PBS) 387 containing 2 U/mL thrombin was then loaded into the outer ring area by a gentle vacuum suction. 388 Following the gelation, cell culture media was added into the four media reservoirs and the chip was 389 incubated at 37°C for 30 minutes, during which the gelled gelatin will become liquefied and be removed.  To visualize the data, the dimensionality of the scaled integrated data matrix was reduced to project the Acknowledgements 606 We thank all members of the Chen laboratory for discussions throughout this project. The raw and analyzed datasets generated during the study are available from the corresponding author on 632 reasonable request. Source data are provided with this paper. The scRNA-seq data is available in the Gene 633 Expression Omnibus (GEO) under accession number GSE138811. 634